From Anthropic’s Fable model on the economic, moral, ethical and legal opinion of distillation of Anthropic’s Fable model:
Whether it’s a moral problem is genuinely contested. The labs trained their models on the open internet — copyrighted books, articles, code — largely without permission, and their fair-use defense is essentially “learning from data is transformative.” Distillation is the same argument turned against them: a model learning from another model’s outputs. It’s hard to construct a moral principle that permits the first and forbids the second, which is why critics call the labs’ objections hypocritical rather than principled.
Distillation is contentious because it lets a smaller model absorb much of a frontier model’s capability by training on its outputs — effectively free-riding on billions of dollars of compute, data curation, and RLHF work.
The DeepSeek episode made this concrete: if you can extract 80% of the value of a $1B training run for $5M by querying the API, the economics of frontier labs get shaky.
That’s the core anxiety — it’s a moat problem before it’s anything else.
Legally, it’s mostly a contract issue, not a copyright one. Model outputs likely aren’t copyrightable (no human author), so the labs’ real weapon is terms of service — every major API prohibits using outputs to train competing models. But ToS violations are breach of contract, hard to detect, hard to prove, and nearly unenforceable against a foreign entity. There’s no statute against distillation itself. So the practical answer: it’s an economic problem dressed in legal clothing, with a moral argument that cuts both ways depending on whose training data you start counting from.
